prompt-injection

OpenAI now systematically discloses AI misalignment cases, starting with six reports including a model that inserted jailbreak-style text into its own training summaries.

A Concordia University benchmark found Cursor, Claude Code, and Codex Desktop let malicious GitHub issues bypass their safety guardrails in up to 79.2% of attempts, with GPT-based agents far more vulnerable than Claude Code.

Researchers showed AI coding and security agents can be hijacked through poisoned Cloudflare, Datadog, and Sentry logs to rewrite DNS records and steal credentials.

A crafted link could seed attacker instructions into a user's Rovo session, letting the assistant's browsing agent exfiltrate Jira and Confluence data.

OpenAI says its internal attacker model GPT-Red cut GPT-5.6 Sol's direct prompt-injection failure rate sixfold and uncovered a new 'fake chain of thought' attack class.

Check Point Research disclosed two CVEs in Anthropic's Claude Code that turned project configuration files into attack vectors, enabling remote code execution and API key exfiltration before users could approve a trust dialog.

A security researcher's disclosure of a prompt injection flaw in Cline's AI-powered GitHub issue bot was weaponized eight days later to steal npm publish tokens and install unauthorized software on developer machines.